JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO.

X, XXXX 2016 1

Cost-Effective Direction of Arrival Estimation using

Pairwise Antenna Geometry
Farhan Rasheed, Noor Muhammed Khan
ACME Center for Research in Wireless Communications (ARWiC)
Capital University of Science and Technology, Islamabad, Pakistan.

Abstract—Cost effective and accurate hardware implementa- coverage and flexibility to receive all kinds of transmissions
tion of Direction of Arrival (DOA) estimation is a challenging with mostly single source targets. Furthermore, all this is
task. In this paper, an endeavor is made to devise a cost required to be packed in a cost effective mobile package with
effective optimized implementation of popular algorithms without
any loss in performance for single source. A novel method to minimum power consumption, hence the need for optimized
distribute any arbitrary antenna geometry into switched pairs implementation techniques.
with common reference element is proposed. This results in Research work on DOA estimation dates back to almost
substantial reduction of material as well as of computational a century where Belinni and Tosi [4], along with Marconi
cost in direction finding. Comparative analysis of the results [5], attempted to use directional antenna elements to perform
achieved in the proposed method with those exhibited in the
standard implementations of any algorithm indicates substantial direction finding. Then focus shifted towards use of multiple
reduction in complexity and cost while improving performance. antennas for direction finding as proposed by Adcock [6],
This technique is applicable particularly in electronic warfare which has now become a common term for direction finding
where targets are mostly single sources. systems. First attempt to automatically estimate the locations
Index Terms—DOA, UCA, ULA, FPGA, Correlation, MUSIC. of emitters using sensor arrays was presented in 1950 by
Bartlett [7]. The method applied classical spectral Fourier
analysis to spatial analysis. In [8], Capon presented his famous
I. I NTRODUCTION method. It relies on the idea of constraining the gain of the
array to unity in a given direction while minimizing power
D OA estimation is a key component in modern commu-
nication systems. In [1], authors have compiled decades
of research on DOA estimation and elaborated strengths and
in other directions. Later with the advent of computer age
complex high resolution direction finding methods based on
weaknesses of various methods. A typical DOA estimation matrix decomposition into signal and noise subspace started to
problem with three signals arriving from φ1 , φ2 , and φ3 appear. Most common of which are MUSIC [2], by Schmidt
impinging upon seven elements displaced by distance d is and ESPIRIT [3], by Kailath followed by their numerous vari-
depicted in Fig. 1. In commercial applications, we are usually ants. Previously mentioned methods are now termed as classic
DOA methods while subspace based methods like MUSIC and
ESPIRIT are usually termed as modern DOA methods. MU-
SIC became more popular due to its ability to be used on any
arbitrary antenna geometry and reliability of estimation results.
Its popular variants include Smooth MUSIC proposed by Qi,
Chen, Wang and Zhan [9], and Spatially Selective MUSIC (S2-
MUSIC) proposed by Khallaayoun and Huang [10]. Similarly,
several variants of ESPIRIT have also been proposed over last
three decades. All of the modern estimation methods including
Fig. 1. DOA estimation problem with 3 incident signals MUSIC and ESPIRIT are adaptive algorithms as they are
dependent on an estimate of the correlation matrix. Estimation
restricted to particular frequency bands allowing easy and of this matrix is a computationally complex as several samples
cost effective receiver design with more number of channels. are required to estimate it with the assumption that all samples
This allows application of higher order DOA techniques like follow the same statistics which may not be valid particularly
MUSIC [2], and ESPIRIT [3], which benefit from more in dynamically changing RF environments. As a result non-
number of channels. However, in military and law enforce- statistical or direct data domain technique known as Matrix
ment applications, the requirements include wide frequency Pencil [11], was proposed by Hua and Sarkar. Matrix Pencil
was originally developed for the estimation of the poles of a
Farhan Rasheed is with the Department of Electrical Engineering, Capital
University of Science and Technology, Islamabad, Pakistan e-mail: farhan- system. However, it can be applied as well to DOA estimation
[email protected]. for uniform linear arrays because such an array behaves similar
Prof Dr Noor Muhammad Khan is with the Department of Electrical Engi- to a filter system.
neering, Capital University of Science and Technology, Islamabad, Pakistan
e-mail: [email protected]. Implementation/verification of research in DOA estimaton
Manuscript received XXXX XX, 2016; revised XXXX XX, 2016. poses serious issues when actually realized in hardware due
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 2

to various factors including antenna/hardware imperfections, MUSIC and Smooth MUSIC are discussed. In section 3, the
computing resources required and cost considerations etc. proposed method is described alongwith its hardware imple-
Lot of research has already been done by scholars on op- mentation in detail. In section 4 performance of the proposed
timized hardware implementation of direction finding algo- method with that exhibited by the standard implementations
rithms, however, most of the work has been focused on of the correlation and MUSIC algorithms with same number
optimum mathematical implementation on DSP [12], or on of antenna elements is compared. Various performance degra-
FPGAs [13] - [21]. In these articles, authors have presented dation effects due to the proposed implementation are also
optimum methods to solve matrix decomposition problem discussed. In the end, conclusions and directions for future
which appears in all subspace and most of the non subspace work are presented in section 5.
methods. However, as the number of antennas increases matrix
decomposition complexity increases exponentially. Authors in II. S YSTEM M ODEL
[14] [15] have used QR factorization to solve eigen decompo- The best accuracy possible in DOA estimation can be
sition in MUSIC and ESPIRIT. In [17], authors have solved calculated using Cramer Rao Bound (CRB) which sets the
the matrix decomposition problem in Root MUSIC by using upper limit for accuracy for a given SNR. In [25], author has
Singular Value Decomposition (SVD) algorithm and Parallel derived CRB for DOA estimate for a received signal vector
Polynomial Rooting utilizing FPGA platform. Similarly in with the average amplitude α corrupted by white Gaussian
[21], authors have exploited parallelism available in Jacobi noise with power σ 2 on antenna array with N elements
Method again on FPGA to ensure realtime solution of direction separated by distance d given by Eqn. (1).
finding problem. In Jacobi method matrices are solved in 6σ 2
parallel on any parallel processing platform like FPGA by var(φ) ≥ (1)
|α|2 N (N 2 − 1)(kd)2 sin2 (φ)
breaking input correlation matrix into smaller matrices of
lower order and then solving them in parallel to achieve The signal can be modeled as M signals from M different
higher speeds. As far as exploiting the antenna structure is sources incident on an antenna array. These signals also
concerned, Smooth MUSIC, splits the antenna into smaller contain noise from all sources including receivers own thermal
subarrays but all signals must belong to same time instant noise. However, for our research we will only consider thermal
therefore still requiring same number of receivers. Aim here noise of the receive channels. This signal can be represented
is to incorporate spatial smoothing for the DOA estimation of by Eqn. (2),
M
coherent sources and not the hardware optimization. Similarly
X
x= am s(φm ) + n (2)
in S2-MUSIC, antenna is divided into overlapping subarrays m=1
forming a beamformer in the general direction of the source
while our aim is to estimate φm , where m = 1, ...M .
only improving performance while reducing computational
load however, it requires detection of general direction of the A. Correlation Method
source prior to the application of the algorithm in the first
Correlation is probably the most simplest technique em-
stage.
ployed for DOA estimation problem. This method has been
The existing methods mentioned in the literature review
selected due to its simplistic approach and excellent perfor-
require several receive channels usually equal to the number
mance in single user scenario. Cauchy-Schwarz inequality
of antenna elements in the hardware making it costly and
states that, as a function of φ, sH (φ)s(φm ) has a maximum
complex to implement. In this paper, we endeavor to find a so-
at φ = φm . Therefore, using the correlation method, we can
lution right at the antenna stage alongwith some modifications
calculate pseudo power spectrum Pcorr (φ) for each value of
in calculation of the estimation results, to reduce complexity
φ where,
and cost for all subsequent receiver and baseband stages.
Pcorr (φ) = sH (φ)x (3)
The proposed solution uses switching of antenna elements
with respect to a reference element to convert any antenna Pcorr (φ) is a non-adaptive estimate of the spectrum of the
geometry into smaller uniform linear arrays composed of incoming data. We can obtain DOA estimates by finding the
only 2 elements each then estimating DOA using standard M largest peaks of this plot for M signals impinging on our
algorithms and in the end results from each array are summed array. Moreover, if our antenna array is a Uniform Linear
to get the overall DOA estimation result of the antenna. This Array (ULA) with equal distance between its elements then
will reduce the order of the correlation and decomposition the steering vector s(φ) is equivalent to Fourier coefficients,
matrix operations to 2x2 making it easier to process in i.e., the correlation result obtained from Eqn. (2) is equivalent
hardware. Furthermore only 2 channel receiver is required for to a Discrete Fourier Transform of the received signal vector
its hardware implementation reducing the cost substantially x. It can be proved that in case where the impinging signal is
while also reducing power consumption, size and complexity from a single user only, correlation technique is optimal (in
of direction finding hardware. This technique also improves the maximum likelihood sense) which makes this technique
DOA estimation due to spatial smoothing effect of overlapping particularly suitable as compared to all other techniques in
sub arrays. single user situation.
The rest of the paper is organized as follows: In section 2, Results obtained from correlation method for a signal im-
popular DOA estimation algorithms both classic e.g. correla- pinging on a 5 element Uniform Circular Array (UCA) is
tion method as well as modern subspace based algorithms e.g. shown in figure below,
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 3

1
In contrast to simple correlation method described earlier,
0.9
MUSIC provides very distinct peaks on possible angles of
0.8 arrivals. This helps in better resolution of closely spaced sig-
Normalized Power Spectrum
0.7 nals which is otherwise difficult in simple correlation method.
0.6 Following figure shows typical results of MUSIC for a signal
0.5
arriving from -140 degrees and impinging on a 5 element
0.4
UCA,
0.3

0.2

1
0.1

0 0.9
-200 -150 -100 -50 0 50 100 150 200
DOA (degrees) 0.8

Normalized Power Spectrum

0.7

Fig. 2. Typical Correlation Method Pseudo-Spectrum for a signal arriving 0.6

from -140 degrees 0.5

0.4

0.3
B. MUSIC: Multiple Signal Classification 0.2

MUSIC is probably the most popular technique employed 0.1

for DOA estimation problem due to its flexibility, statistically 0

-200 -150 -100 -50 0 50 100 150 200
DOA (degrees)
robust results and ability for its application on arbitrary an-
tenna geometries [1]. It is due to these properties that we have
Fig. 3. Typical MUSIC Method Pseudo-Spectrum for a signal arriving from
selected MUSIC for our proposed hardware implementation. -140 degrees
MUSIC is dependent on the correlation matrix of the
received data like any other adaptive technique. Using the data
model in Eqn. (2). Assuming that the different signals to be C. Smooth MUSIC
uncorrelated, the correlation matrix of x can be written as,
Over the decades several avenues of improvement have been
R = E[xxH ] (4) investigated in the original MUSIC algorithm as a result there
are several variants of the MUSIC algorithm available now,
which reduces to, including Cyclic-MUSIC and Smooth-MUSIC. However par-
R = Rs + σ 2 I (5) ticular focus has been on improvements in resolving correlated
where Rs is the signal correlation matrix while σ 2 is the signals due to this Smooth-MUSIC is particularly interesting
noise power multiplied with identity matrix to form noise because it overcomes the MUSIC assumption that all incoming
correlation matrix. Hence it can be decomposed into signal signals are uncorrelated. In a communication situation, assum-
and noise matrices. This implies that eigen vectors of R is ing flat fading, there may be multipath components from many
clearly composed of signal eigen vectors and noise eigen directions. These components would be correlated with each
vectors. R is a N xN matrix with rank M and as noise is other. Correlated components reduce the rank of the signal
uncorrelated with signal therefore R has N − M noise eigen correlation matrix Rs , resulting in more than (N − M ) noise
vectors corresponding to zero eigen values which can be called eigenvalues.
Qn . Therefore it can be shown that these noise eigen vectors Inorder to mitigate effects of correlated components
are orthogonal to M steering vectors. This forms the basis of Smooth-MUSIC divides N antenna elements into L overlap-
MUSIC algorithm. If Qn is N x (N − M ) matrix of noise ping subarrays, composed of P elements each. For example,
eigen vectors then MUSIC pseudo-spectrum can be written as, elements 0 through P − 1 would form subarray 0 while
elements 1 through P would form subarray 1 and so on.
1 Therefore the resultant number of subarrays L would be given
PM U SIC (φ) = (6)
kQH
n s(φ)k
2 by,
In practical situations the original correlation matrix R is not L=N −P +1 (8)
available therefore it is estimated based upon the received data. Now data obtained from each subarray is used to compute
This estimation requires averaging over several snapshots of L correlation matrices, where each correlation matrix has
data. The number data snapshots used is called the manifold dimension P P . The standard MUSIC algorithm is then used
size. The estimated correlation matrix R is calculated in to estimate directions of arrival from the smoothed correlation
Eqn.(7), matrix,
K L−1
1 X 1 X
R= xk xH
k (7) RL = Rl (9)
K L
k=1 l=0

where xk is the k-th snapshot. In [22], the authors prove that Above manipulation makes the signal correlation matrix com-
one requires at least K > 2N so that the SNR is within 3dB ponent of RL full rank again. As a result, Smooth MUSIC can
of the optimum. estimate the DOA of up to L − 1 correlated signals.
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 4

III. P ROPOSED H ARDWARE I MPLEMENTATION of target signal only for DOA estimation process as shown in
As highlighted in the introduction, present research on opti- Fig 5.
mized hardware implementation of DOA estimation methods
is primarily focused on reducing computational cost after
reception of signal vector for a given number of antenna
elements. For N element antenna system receive vector x for
k-th sample can be represented as,
x(k) = [x0 (k), x1 (k)....xN (k)]T (10)
Our method break this receive vector into sub vectors using
RF switch at the antenna into N-1 pair where one element is
common to each pair and it will be called reference element.
Thus N-1 new vectors are formed as follows,
x1 (k) = [x0 (k), x1 (k)]T (11)
T Fig. 5. Low Pass Filtering before DOA processing
x2 (k) = [x0 (k), x2 (k)] (12)
..
. (13)
xN −1 (k) = [x0 (k), xN −1 (k)]T (14) A. Hardware Description
Here similarity with Smooth MUSIC is evident which also A custom hardware was designed based upon two channel
divides antenna array into overlapping subarrays. However, Software Defined Radio (SDR) for testing of our proposed
unlike Smooth MUSIC, our method restricts the number of method in real world scenario. Fig 6 shows our complete
antenna in each array to two due to which we can distribute implementation however, we will restrict our discussion to
any arbitrary antenna system into smaller 2 element ULAs. direction finding hardware only for this study,
Sub division of our UCA antenna into four subarrays is shown
in Fig 4, Now receive vector from each pair as shown in Eqn

Fig. 6. Overview of hardware implementation

1) Antenna Design: Our antenna system consists of 5

Fig. 4. 5 element UCA antenna array divided into 4 subarrays element dipole (tuned at 150MHz) based uniform circular
array custom designed specifically for our implementation. It
11 will be used to obtain a 2 x 2 correlation matrix R as shown consists of antenna elements and a frame. Fig 7 shows the
in Eqn 7. This data will now be used in our standard algorithms actual antenna element,
already discussed. Since each receive vector has dimension The frame houses RF switch as well as various sensors
2 therefore only two channel receiver is required can be including magnetometer, gyro and accelerometer. The antenna
shared in time for each subarray. More importantly, since R frame is shown in Fig 8,
is 2 x 2 matrix to its eigen decomposition is very easy and 2) RF Block: RF block consists of RF switch and RF
computationally efficient despite the fact that it has to be done frontend. RF switch comprises of AS204 wideband single
N − 1 times. Results produced by this method due to spatial pole four throw (SP4T) bidirectional RF switching IC from
smoothing and sampling of signal at different angles. However Skyworks having typical switching time of 75 ns and provides
the proposed method is different from Smooth MUSIC which a continuous coverage from 300 KHz to 4.4 GHz. Actual
has none of the elements common in all subarrays and there antenna switch hardware is shown in Fig 9,
is no limit of subarray size as a result requiring more number The frontend is based upon AD-FMCOMMS3-EBZ eval-
of receivers. Moreover, due to 2 x 2 correlation matrix, our uation module from Analog Devices providing fully syn-
proposed method can be used only for single signal source chronized 2 channel wideband RF receivers. Frontend IC is
which is the main drawback. But military radio direction AD9361 which is a highly integrated state of the art 2 channel
finders mostly have single emitters as their targets therefore MIMO cap RF frontend which provides continuous coverage
this drawback does not effect them much. Inorder to ensure from 50MHz to 6 GHz with exceptional noise fig (NF) of
that signal from a single source is processed, a narrow band 2.4 dB and baseband sampling rates of upto 61.44MSps. AD-
lowpass FIR filter is used after RF frontend to ensure selection FMCOMMS3-EBZ evaluation module is shown in Fig 10.
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 5

Fig. 9. RF antenna switching module

Fig. 7. Dipole element tuned at 150 MHz

Fig. 10. RF frontend used for hardware implementation

B. Hardware Implementation Issues

The main problem experienced during any hardware imple-
mentation of direction finding is phase synchronization. In our
proposed implementation this issue is further accentuated due
to presence of RF switch in the receive path. Major problems
experienced during implementation and their remedial mea-
sures undertaken are discussed in this section.
1) Antenna Imperfections: Antenna imperfection is a com-
Fig. 8. Antenna frame with radius of 0.48m mon problem in DOA estimation and our antenna being de-
signed and manufactured indigenously was particularly prone
to it. Effects due to imperfections of the antenna were miti-
3) DF Block: DF block consists of DF processor and gated by obtaining the actual steering vector of our antenna
DF Software modules. Zedboard has been employed for DF instead of estimated one. For calculation of steering vector
processing which is based upon Zynq 7 series FPGA platform a 100m radius circle was marked with one degree steps in
from Xilinx. This platform provides processing resources a open ground. A CW emitter was placed on each marked
based upon dual core ARM Cortex-A9 processor with NEON degree and phase differences induced be each element were
floating point extensions as well as programmable logic (PL) obtained. Thus all the effects of imperfection were part of the
FPGA for implementing high speed algorithms in a sin- steering vector therefore there effects were nullified.
gle package. Presently Filters and Digital Down Converters 2) Mutual Coupling Amongst Antenna Elements: Another
(DDCs) are implemented in PL while DF algorithm is im- major issue faced during practical implementation of di-
plemented in software running on ARM processor. It runs a rection finding algorithm is the effect of mutual coupling
standard linux operating system with necessary drivers for RF by antenna elements [23]. Each antenna element receiving
frontend facilitating software development. Zedboard is shown electromagnetic waves re-radiates some of its energy back
in Fig 11, to the surroundings. Amount of energy re-radiated depends
Standard driver from Analog Devices for linux platform upon several factors including impedance matching between
are provided with AD-FMCOMMS3 eval board used in our antenna and the receiver cable. Mutual coupling effect is more
hardware implementation. These were modified and direction pronounced if multiple antenna elements are placed close
finding software was then built upon these drivers. Proper and together otherwise if sufficient distance > λ/2 exists between
convenient testing of the direction finding hardware in the field the antenna elements this effect is very less.
required a Geographic Information System (GIS) software Mutual coupling effects can be compensated by using
inorder to view shoots taken by the direction finder on map. A compensation matrix which can be calculated in several ways
complete GIS solution was developed for this purpose using depending upon the type of antenna e.g. transmitting antenna
opensource libraries for map manipulation. Screenshot from or the receiving antenna as effects are different in both cases.
this software during actual field is shown in Fig 12, Similarly it also depends upon DOA of the received signal. If
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 6

Fig. 11. Xilinx ZedBoard used for DOA processing

Fig. 13. RF switch PCB design

IV. R ESULTS D ESCRIPTION

The proposed method was simulated in MATLAB and was
also implemented on actual hardware while the conventional
methods were only simulated in MATLAB. To be fair main
comparison will be drawn between the simulated results only
under same test conditions, steering vector and received data
while performance on actual hardware will be presented sep-
Fig. 12. Screenshot of DOA estimate plot on actual map arately. It is important to note that during simulation effect
of RF switch in case of proposed method was also included
in simulations to make results more realistic. Moreover, the
s(φ) is the steering vector of array without mutual coupling test signal was amplitude frequency and phase modulated for
and C is the mutual coupling compensation matrix then different tests.
compensated steering vector sc (φ) is given as follows,
sc (φ) = Cs(φ) (15) A. Effect of Splitting Antenna into Pairs
It is important to understand what is the actual effect of
In our implementation the steering vector is already com- distributing the antenna into pairs because each pair becomes
pensated because we have calculated it from actual antenna a ULA antenna with two elements. These ULA subarrays
response from all angles instead of simulated response. provide two DOA estimates each using any algorithm capable
3) Parasitic Effect of Antenna Elements: In a antenna of operating on 2 element ULA. However, if all N - 1 results
theory, a passive radiator or parasitic element is a conductive are added, out of the two estimates only the one which are
element, which is not electrically connected to anything else from actual DOA align and reinforce each other while the
[24]. These elements cause maximum re-radiation of the other ones do not align and therefore dissipate as shown in Fig.
received signals significantly effecting array response. Since 14 and 15 for our 5 element UCA antenna, Similar effect can
DOA estimation is totally dependent upon the array response
therefore parasitic elements will effect DOA estimation pro-
cess significantly. Our antenna uses RF switching for selection
of elements therefore unconnected elements would become
parasitic elements effecting our results. To overcome this issue,
we selected a non-reflective RF switch for our design where
unconnected elements are terminated across 50 Ohm load
inside the switch.
4) Effects of RF Switch: As mentioned above our proposed
design also incorporates an RF switch just after the antenna
system. This switch is not in the path of reference element
therefore phase response of switch only effects the other
channels disturbing phase synchronization. To mitigate this
additional capacitors were placed in the reference path on Fig. 14. Correlation Spectrum generated from each pair
the same PCB as the switch. The values of capacitors were
chosen in a way that effect of RF switch capacitance became be seen for implementation of MUSIC with proposed method,
negligible. RF Switch PCB is shown in Fig. 13, This effect is the basic reason that our method can be used
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 7

MUSIC UCA
1

Normalized Power Spectrum

0.8

0.6

0.4

0.2

0
-200 -150 -100 -50 0 50 100 150 200
DOA (degrees)

MUSIC UCA Proposed

1

Normalized Power Spectrum

0.8

Fig. 15. MUSIC spectrum generated from each pair 0.6

0.4

0.2
with any arbitrary antenna system.
0
-200 -150 -100 -50 0 50 100 150 200
DOA (degrees)
B. Comparison with Standard Correlation Method
In this section simulation results based upon standard and
Fig. 17. Comparison with standard MUSIC Method
proposed switched implementation of correlation method have
been presented. Fig. 16 shows results for a signal received
at 20dB SNR from -140 degrees, We can see here that the
a result of splitting antenna into ULAs. However, additional
peaks will always be less then the actual peak as described in
1
Correlation UCA section (4.2). Again since we know that the source is single
to we use the largest peak only for DOA estimation.
Normalized Power Spectrum

0.8

0.6

0.4
D. SNR Measurements
0.2

Here we present the performance comparison of standard

0
-200 -150 -100 -50 0
DOA (degrees)
50 100 150 200 implementation with proposed implementation method based
upon root mean square error vs SNR. Fig. 18 shows the
Correlation UCA Proposed
1
simulation results based upon 100 trials with different random
0.9
Normalized Power Spectrum

data for each trial, Result show that the best performance
0.8

0.7

0.6 RMS Error Plot

80
0.5 Correlation UCA
70 Correlation UCA Proposed
0.4 MUSIC UCA
MUSIC UCA Proposed
0.3 60
-200 -150 -100 -50 0 50 100 150 200
DOA (degrees)
RMS Error (degrees)

50

40

Fig. 16. Comparison with standard Correlation Method 30

20
conventional correlation method provides one distinct peak
only while additional lobes are appearing in proposed method 10

as a result of splitting antenna into ULAs. However, additional 0

0 5 10 15 20 25 30

peaks will always be less then the actual peak as described in SNR (dB)

section (4.2). Since we know that the source is single to we

Fig. 18. Mean Square Error vs SNR Plot for 100 trials (K=51)
use the largest peak only for DOA estimation.
is offered by the proposed implementation with correlation
C. Comparison with Standard MUSIC method particularly at low SNRs. This is understandable be-
In this section simulation results based upon standard and cause correlation method is optimal for single source detection
proposed switched implementation of MUSIC method have as discussed in section 2. The second best result is given
been presented. Fig. 17 shows results for a signal recieved at by MUSIC with proposed implementation method. While
20dB SNR from -140 degrees, Again we can see here that the standard implementation struggle under the same simulation
conventional MUSIC method provides one distinct peak only conditions. These results are as a result of spatial smoothing
while additional lobes are appearing in proposed method as effect due to splitting of antenna into overlapping subarrays.
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 8

E. Estimation Error a two channel transceiver implementation for typical WLAN

In this section distribution of measured DOAs is presented implementation supporting costs few tens of dollars while a
to understand the underlying pattern. Results are presented high performance wideband wide frequency range two channel
for a signal with 5dB SNR with 100 trials in Fig. 19, Again receiver system like the one used in our implementation costs
hundreds of dollars.
Histogram Correlation Histogram Correlation Proposed
Another major advantage is reduction in hardware com-
30 30
plexity and size while still improving upon performance. One
20 20 of the major issues in more receive channels is the clock
No of Hits

No of Hits
10 10
generation and distribution system. It is important to note
that 2 channel single chip receivers similar to one we have
0
-160 -140 -120
0
-160 -140 -120 used are easily available where clock and local oscillator
DOA (degrees) DOA (degrees)

30
Histogram MUSIC Histogram MUSIC Proposed
30
synchronization is already ensured on the chip making receiver
design extremely easy.
20 20
Impact on accuracy and precision of hardware is also worth
No of Hits

No of Hits

10 10 mentioning. It is a direct consequence of reduction in both

hardware and software complexity. Our proposed implemen-
0 0
-160 -140
DOA (degrees)
-120 -160 -140
DOA (degrees)
-120
tation is simple which minimize the unforeseen errors and
probability of hardware induced errors common in complex
Fig. 19. DOA estimation (5dB SNR) Histogram Comparison for 100 trials hardware.
(K=51) Our case of 5 element UCA using conventional imple-
mentation requires solution of a 5x5 matrix which more
the trend continues and we find that the closest distribution is compute intensive as compared to solution of five 2x2 ma-
offered by correlation method with proposed implementation trices. Therefore both memory and processing requirements
followed by MUSIC with proposed implementation. If we reduce significantly. Similarly, greatest gains would come from
compare the standard deviation of the estimates produced by implementation of proposed method on FPGA where required
each method for results in Fig 19, the improvement becomes number of gates and other resources would significantly re-
more evident as shown in table I, duce.
Ser SNR Type Standard Deviation (K=51) In terms of overall computation time our proposed solution
1 5dB Correlation 8.16 shows a little disadvantage particularly in large number of an-
2 5dB Correlation Proposed 2.3 tenna elements. Although reduced computational complexity
3 5dB MUSIC 7.67
4 5dB MUSIC Proposed 3.7 already highlighted reduces computational time of software,
TABLE I however, switching time of RF switch adds to total compu-
S TANDARD D EVIATION CHART FOR THE RESULTS IN F IG 4.7 tational time required for a single DOA estimate. In general
overall compute time is slightly increased as compared to stan-
dard implementations, however, the effect can be minimized
by using higher speed RF switches. Table III shows typical
F. Performance on Actual Hardware execution times for QR Decomposition required for calculation
Hardware implementation was tested in the field and of eigenvectors on an FPGA for different matrix sized [14].
demonstrated excellent results. All tests were conducted in Important here is to note that impact on time as matrix size
environment with scatted trees and multi story buildings. DOA increases.
estimation results on various SNR levels, signal types and
Ser M N Cycles Time (µs)
angles are tabulated in table II. Each entry consists of five 1 3 3 939 3.76
measurements and result of each measurement is shown, 2 5 5 2,795 11.18
3 7 7 6,027 24.11
Ser SNR Type Actual DOA Estimated DOAs (k=32) 4 9 9 10,971 43.88
1 30 Digital 40 40,40,41,40,41 5 10 10 14,190 56.76
2 30 Digital 80 81,81,80,80,79 TABLE III
3 30 Analog 181 181,181,181,181,181 T YPICAL EXECUTION TIME FOR FPGA (C LOCK 250MH Z ) QR
4 20 Digital 119 120,120,120,121,120 D ECOMPOSITION IMPLEMENTATION FOR AN M X N MATRIX
5 10 Analog 119 120,121,120,120,121
TABLE II
DF RESULTS OF PROPOSED IMPLEMENTATION OF CORRELATION METHOD

V. C ONCLUSION
DOA estimation is a subject of continuous research due to
G. Hardware Optimizations Achieved its pivotal role in modern high speed digital communications
The most important and desired effect is reduction in cost of and electronic warfare applications. Latest techniques like
hardware. For a 5 element system which would normally re- co-prime arrays and direct data domain solutions keep on
quire 5 channel synchronous receiver, cost estimate is reduced emerging in times to come. This indicates significant poten-
to 60 percent and for 9 elements to 20 percent. For reference tial for further research in this field. Nevertheless, testing
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 9

such research in the field requires complex and extremely [2] R. O. Schmidt, Multiple emitter location and signal parameter estimation,
costly hardware. That is why implementation of significant IEEE Transactions on Antennas and Propagation, vol. 34, no. 3, pp. 276
- 280, 1986.
part of the research work remains on simulated platforms. [3] R. Roy and T. Kailath, ESPRIT-Estimation of signal parameters via
Therefore, optimization in hardware implementation remains rotational invariance techniques, IEEE Transactions on Acoustics,
an important goal particularly for practical applications like Speech, and Signal Processing, vol. 37, no. 7, pp. 984 - 995, 1989.
[4] Bellini, E. and Tosi, A directive system of wireless telegraphy, Elect.Eng.
electronic warfare where factors like cost, size, complexity, 2 (11), pp. 771-775. 1906.
power requirements and portability always exert significant [5] Marconi, G, On Methods Whereby the Radiation of Electric Waves May
weight. be Mainly Confined to Certain Directions, and Whereby the Receptivity of
a Receiver May be Restricted to Electric Waves Emanating from Certain
In our dissertation, we have made an effort to optimize Directions, Proceedings of the Royal Society of London. Series A,
implementation of already available DOA estimation methods Containing Papers of a Mathematical and Physical Character, vol. 77, no.
by splitting any arbitrary antenna geometry into multiples of 518, pp. 413-421, April 30, 1906,
[6] Adcock, F, British Patent Specification No.130490 1919.
2 element ULA antennas. Our proposed implementation is not [7] Bartlett, M. S, Periodogram Analysis and Continuous Spectra,
only significantly less costly but also easier to implement while Biometrika vol. 37, pp. 1-16, 1950.
offering excellent results comparable with those exhibited in [8] J. Capon, High-Resolution Frequency-Wavenumber Spectrum Analysis,
Proceedings of the IEEE, vol. 57, no. 8, Aug 1969.
standard implementation methods. Complexity mainly arises [9] Chongying Qi, Zhijie Chen, Yongliang Wang and Yongshun Zhan, DOA
from the calculation and solution of correlation matrices of Estimation for Coherent Sources in Unknown Nonuniform Noise Fields,
the received vector. Reduction in correlation matrix has been IEEE Transactions on Aerospace and Electronic Systems, 2007.
[10] Khallaayoun, A., and Huang, Y, Spatial Selective MUSIC for Direction
proposed in several algorithms for reducing complexity in ma- of Arrival Estimation with Uniform Circular Array, IEEE Antennas
trix decomposition particularly in sub-space methods however, and Propagation Society Annual Symposium, July 2007.
our approach not only reduces the required receive channels [11] Y. Hua and T. Sarkar, Matrix pencil method for estimating parameters of
exponentially damped/undamped sinusoids in noise, IEEE Transactions
to two but also limits the size of the correlation matrix to on Acoustics, Speech and Signal Processing, vol. 38, pp. 814-824, May
2x2; hence, optimizing both hardware as well as software 1990.
implementation. The results shown in our simulations and on [12] S. Dikmese, A. Kavak, K. Kucuk, S. Sahin, A. Tangel, H. Dincer, Digital
signal processor against field programmable gate array implementations
physical hardware validate the authenticity of our proposed of space-code correlator beamformer for smart antennas, IET Mi-
implementation method. This makes our work suitable for use crowave. Antennas Propagation, vol. 4, issue. 5, pp. 593 - 599, 2010.
in actual direction finding applications. [13] H. Arai and K. Ichige, Hardware Implementation of Smart Antenna
Systems for High Speed Wireless Communication, International Union
The future research potential that exists in our proposed of Radio Science, Proc. Of Gernal Assemblies, paper ID 01157, 2005.
research is four fold. As far as the implementation of the [14] C. Dick, F. Harris, M. Pajic, and D. Vuletic, Implementing a Real-Time
proposed method is concerned, future research can be done Beamformer on an FPGA Platform, Xcell Journal, pp. 36-40, 2nd
Quarter, 2007.
to investigate its performance in 2D DOA estimation. Com- [15] M. Kim, K. Ichige, and H. Arai, Implementation of FPGA based Fast
parisons can be drawn with standard 2D DOA estimator imple- DOA Estimator using Unitary MUSIC Algorithm, Vehicular Technology
mentations on performance, cost effectiveness, computational Conference, vol. 1, pp. 213-217, Oct 6-9, 2003.
[16] M. Kim, K. Ichige, and H. Arai, Real-time Smart Antenna System
complexity and execution speed etc. Incorporating FPGA-based Fast DOA Estimator, Vehicular Technology
Furthermore, testing performance of DOA estimation algo- Conference, vol. 1, pp. 160-164, Sept 26-29, 2004.
rithms other then the ones used in our research using the [17] Mrudula V. Athi, High Performance, Low Cost Subspace Decomposition
And Polynomial Rooting For Real Time Direction Of Arrival Estimation:
proposed implementation method is also a potential avenue Analysis And Implementation, Michigan Technological University, 2014.
for future research. Changes in behavior and performance of [18] Nizar Tayem, Real Time Implementation for DOA Estimation Methods
other algorithms due to modification in antenna geometry is on NI-PXI Platform, Progress In Electromagnetics Research B, vol. 59,
pp. 103-121, 2014.
of particular interest. [19] S. Dikmese, A. Kavak, K. Kucuk, S. Sahin and A Tangel, FPGA
Similarly on hardware implementation side, our present Based Implementation and Comparison of Beamformers for CDMA2000,
work in particular the direction finding algorithm is done on Wireless Pers Commun, DOI 10.1007/s11277-009-9855-42014, Springer,
Nov, 2009.
ARM Cortex A9 processor available in Zynq platform used in [20] Sachin Khedekar, Rachna Kumari, Sitakanta Maharatha and Dr. Mainak
our hardware. However, our proposed solution can be ported Mukhopadhyay, Implementation of a Proposed 2D DOA Estimation
to FPGA part also available in Zynq platform and other hybrid Algorithm on an FPGA Platform, International Journal of Current
Engineering and Scientific Research (IJCESR), vol-1, Issue-2, 2014.
compute platforms where various performance parameters par- [21] Minseok Kim, Koichi Ichige and Hiroyuki Arai, Design of Jacobi EVD
ticularly cost and execution times can be examined. Porting of processor based on CORDIC for DOA estimation with MUSIC algorithm,
our work to FPGA will definitely improve the direction finding IEICE Transactions on Communications vol.E85-B no.12 pp.2648-2655,
Dec. 2002.
hardware further making it into a complete field deployable [22] I. S. Reed, J. Mallett, and L. Brennan, Rapid convergence rate in
product. adaptive arrays, IEEE Transactions on Aerospace and Electronic
Lastly, further research is also recommended in application Systems, vol. 10, No. 6, pp. 853 - 863, Nov, 1974.
[23] S. Dikmese, A. Kavak, K. Kucuk, S. Sahin, A. Tangel and H. Dincer,
of proposed method on large antenna arrays with more than Mutual Coupling in Antenna Arrays, International Journal of Antennas
ten elements. Effects on performance and execution times can and Propagation, vol. 2010, Special Issue, pp. 593 - 599, 2010.
be investigated. [24] Mofolo Mofolo, Albert Lysko and Willem Clarke, A Method of Elec-
tronic Beam Steering for Circular Switched Parasitic Dipole Arrays, IET
Microwave. Antennas Propagation, vol. 4, Issue. 5, pp. 593 - 599, 2010.
R EFERENCES [25] Prof. Raviraj Adve, Direction of Arrival Estimation, Graduate
Lecture Notes, University of Toronto, Canada, Available online at
[1] Krim, H. and M. Viberg, Two decades of array processing research: The ”http://www.comm.utoronto.ca/ rsadve/Notes/DOA.pdf” in April 2016.
parametric approach, IEEE Signal Processing Magazine, vol. 13, no.
4, 67-94. July, 1996.
JOURNAL OF AEROSPACE AND ELECTRONIC SYSTEMS, VOL. XX, NO. X, XXXX 2016 10

Noor Muhammed Khan Biography text here.

PLACE
PHOTO
HERE

Farhan Rasheed Farhan Rasheed has done his

Bachelors in Telecommunication Engineering from
National University of Science and Technology, Pak-
istan and Masters in Electronic Engineering from
PLACE Capitol University of Science and Technology, Pak-
PHOTO istan. His field of research is Electronic Warfare and
HERE Direction of Arrival Estimation.